- cross-posted to:
- technews@radiation.party
- cross-posted to:
- technews@radiation.party
Summary: This was on Linux using kernel mitigations. The performance impact can be very significant to some workloads, like databases and compression, but most users likely will not notice any impacts.
You must log in or register to comment.
This is the best summary I could come up with:
Last week the AMD Inception vulnerability was made public as a speculative side channel attack affecting Zen processors and different mitigation options based on the CPU generation.
There wasn’t too much communication around the performance implications of mitigating Inception while over the past week I have begun benchmarking the software and microcode updates on Ryzen and EPYC processors.
AMD already published the updated Family 19h microcode for EPYC processors in linux-firmware.git while on the consumer/client side AMD partners will be rolling out updated AGESA with the mitigated microcode.
There are though updated patches in development to clean-up this Inception (SRSO) mitigation code and that cleaned up work will likely be upstreamed in the coming days, but ultimately isn’t expected to change the mitigation overhead costs.
For getting an initial idea of the AMD Inception mitigation performance impact, I used an AMD EPYC 7763 (Milan / Zen 3) server running Ubuntu 22.04 LTS and using a custom kernel build as of last week, The following kernel configurations were tested: off - No Inception mitigations.
The “safe RET” mode is the default mode of operation with AMD Zen processors on the Linux kernel versions patched since last week.
The original article contains 369 words, the summary contains 196 words. Saved 47%. I’m a bot and I’m open source!
This is the best summary I could come up with:
Last week the AMD Inception vulnerability was made public as a speculative side channel attack affecting Zen processors and different mitigation options based on the CPU generation.
There wasn’t too much communication around the performance implications of mitigating Inception while over the past week I have begun benchmarking the software and microcode updates on Ryzen and EPYC processors.
AMD already published the updated Family 19h microcode for EPYC processors in linux-firmware.git while on the consumer/client side AMD partners will be rolling out updated AGESA with the mitigated microcode.
There are though updated patches in development to clean-up this Inception (SRSO) mitigation code and that cleaned up work will likely be upstreamed in the coming days, but ultimately isn’t expected to change the mitigation overhead costs.
For getting an initial idea of the AMD Inception mitigation performance impact, I used an AMD EPYC 7763 (Milan / Zen 3) server running Ubuntu 22.04 LTS and using a custom kernel build as of last week, The following kernel configurations were tested: off - No Inception mitigations.
The “safe RET” mode is the default mode of operation with AMD Zen processors on the Linux kernel versions patched since last week.
The original article contains 369 words, the summary contains 196 words. Saved 47%. I’m a bot and I’m open source!
The bot missed the remaining 7 pages and the result of the benchmark:
“Overall it comes down to what workloads you are engaged in whether you may notice any performance difference when upgrading your Linux kernel (or otherwise being patched for Inception on your given OS) on an AMD Zen desktop or server. For the most part users are unlikely to notice anything drastic, aside from some sizable database performance hits in a few cases. It’s unfortunate seeing some of these regressions due to the Inception mitigation but ultimately is unlikely to really change the competitive standing of AMD’s latest wares on Linux. Most of the prior AMD CPU security mitigations have also not resulted in any performance degradation, so this Inception mitigation difference is a bit rare. It also was announced on the same day as Intel Downfall where there was again a sizable hit to Intel CPU performance.”
No difference to -40% hit based on workload. The average seems to be <-10%.